zakariae-boui/onssa-rag-chatbot

Customer-assistance RAG chatbot for the ONSSA website (Morocco) — Streamlit, Ollama (Mistral 7B), FAISS + BM25 hybrid retrieval, French answers grounded in official content. Internship project. # 🇲🇦 Assistant ONSSA — RAG Chatbot > Customer-assistance chatbot for the official **ONSSA** website > (Office National de Sécurité Sanitaire des produits Alimentaires, Morocco — > onssa.gov.ma). It answers visitor questions in **French**, grounded > **exclusively** in ONSSA website content, with source citations — and runs > **100 % locally with no paid API**. Python · Streamlit · Ollama (Mistral 7B) · FAISS · sentence-transformers · hybrid retrieval (vector + BM25) · local voice (Whisper + Piper) --- ## 📑 Table of contents 1. What it does 2. Key features 3. Architecture 4. How a question is answered 5. Evaluation results 6. Technology choices 7. Setup & run 8. Configuration 9. Deliverables map 10. Documentation 11. Project structure --- ## 🎯 What it does A visitor asks a natural-language question about ONSSA (missions, organisation, métiers, réglementation, e-services, contacts, démarches). The system embeds the question, retrieves the most relevant ONSSA text from a local vector database, and a local LLM writes a **grounded French answer with citations**. If nothing relevant is found, it says so instead of inventing — the anti-hallucination rule. | | | |---|---| | **Corpus** | 284 ONSSA pages · 333 sections · **1,187 indexed chunks** | | **Retrieval quality** | **hit@5 = 14/15** (hybrid) vs 13/15 (naive vector-only) | | **Cost** | **€0** — no paid API; everything runs on the local machine | | **Languages** | Answers in French; multilingual embeddings; voice in French | | **Tests** | 24 unit & smoke tests | --- ## ✨ Key features | Area | Feature | |---|---| | **RAG core** | Retrieval **always** precedes generation; relevance gate refuses off-topic questions *before* the LLM runs | | **Hybrid retrieval** | FAISS vector search **+** BM25 keyword search fused with Reciprocal Rank Fusion | | **Grounding** | Strict French system prompt: answer only from extracts, cite sources `[n]`, no legal/medical/veterinary advice | | **Follow-u …